Cleaning up / refactoring Marlin

[UlrichHaboeck75]

Our Marlin code base deserves some cleaning up. We still have left-overs from Marlin's initial AHP which are not necessary due to the coboundary sumcheck, and there are some structural improvements by arkworks we might want to take over.
In short,

• we can get rid of individual degree bounds in the AHP, and rebrand AHP to an algebraic oracle proof instead (to be in alignment with our paper),
• we could move constraint system related traits (i.e. the large part of the ahp/constraint_system.rs file) to ginger (arkworks seems to have done that), and
• we should move the matrix arithmetization to the indexer submodule.

Beyond that, I recommend to switch to the Lagrange kernel instead of the UnnormalizedBivariateLagrangePoly, even if it comes slightly costlier in circuit. This is mainly for the following reasons.

• The matrix arithmetization step becomes much simpler (we simply index the non-zero entries of the plain R1CS matrices ), and
• we are in full alignment with our paper and the security analysis therein.

If we keep with the generalized derivative, I at least opt to remove the double transposition of matrices: The indexer outputs the transpose of the representation with respect to the (non-normalized) kernel (to be in alignment with the lincheck from Fractal), and then again transposes the arguments for Marlin's t-polynomial in the inner sumcheck.

For details see the TODO's I have pushed in a separate branch inline_docu for our cargo doc.

[UlrichHaboeck75]

P.S.: As we need to stay compatible with the mainchain verifier, I'd opt to provide two versions of Coboundary Marlin in our repo: coMarlin, which is in full alignment with our paper, and coMarlin_gd which works with the generalized derivative.